AI AGENTS IN PRODUCTS
AI that runs inside your product.
Altun Core builds LLM integration as a feature in the screen your users already open, not as a separate tool. The model call stays server-side, output is evaluated, and the product keeps working when the model does not answer.
Technologies used
- OpenAI / Gemini API
- LangChain
- Vector database
- Python
- FastAPI
Providers and tools are chosen per project.
What in-product AI means
Four patterns cover most of the work. All four run in the screen a user already uses, not in a separate tool.
- 01Assistant and chat: the user asks inside the product, and the answer comes from the product's own data.
- 02Document and form reading: fields are extracted from an uploaded file and shown back to the user already filled in.
- 03Classification and routing: an incoming record, request or message is labeled and lands on the right screen.
- 04Search over your own content: the passages closest to the question are found in your content first, then passed to the model. That method is called RAG.
All four share the same groundwork; starting with one is usually enough.
How it is built
An AI feature follows the same rules as the rest of the product: version control, review, and a way back.
- 01The model provider sits behind a server-side API; keys never reach the client.
- 02Prompts and the tools the model may call live in the repository, are versioned, and read as a normal diff.
- 03Output is evaluated against a fixed set of sample inputs; the same set runs again every time a prompt changes.
- 04There is a fallback path for when the model does not answer or the answer cannot be used.
- 05Anything irreversible keeps a human review step.
Data and privacy
Which model provider receives your data, where it is stored, and which setting keeps it out of training are stated in writing. If the data must not leave your systems at all, a model running on your own server is considered as an option.
This is not a guarantee against leakage; it is a written record of which data goes where.
Cost and limits
An AI feature is not a fixed-cost screen. Each request is paid for, and the answer is produced at that moment.
- 01Cost follows the number of requests and the amount of text processed; it is estimated during design and measured once live.
- 02Responses take longer than an ordinary query, so the interface is designed to wait for them.
- 03A model can be wrong and convincing at the same time. The source behind an answer is shown, and critical steps ask for confirmation.
- 04No feature goes live before what counts as a correct answer is written down.
Decide where it starts.
Describe the product, the feature you have in mind, and the data available for it. Which pattern fits, what gets evaluated, and where the cost lands are settled in the first call.
Start a project